5
$\begingroup$

The set of SPD matrices, $\mathbb{P}_n := \{X \in \mathbb{R}^{n \times n} | X=X^T, X \succ 0 \} $, forms a differentiable manifold.

Claim: The tangent space at a point, $A, T_A\mathcal{P}_n$ is the space of symmetric matrices.

Info: I have seen this mentioned in a paper (https://www.ncbi.nlm.nih.gov/pubmed/27845666), but (with my limited understanding of differential geometry) can't see how this occurs (although it may actualy be obvious to someone with a more rigorous maths education). I have also checked out the book they referenced when making this claim (http://www.cmat.edu.uy/~lessa/tesis/Positive%20Definite%20Matrices.pdf) in order to understand it. Sadly the section that I think they are referencing (Start of chapter 6) is far too formal for me. For completeness I will write out what is written there, and if it is the proof I am looking for then an explanation of some of the jumps in the steps would answer my question.

1) The space $\mathbb{M}_n$ is a Hilbert space with inner product $ \langle A,B \rangle = tr A^*B $ and the associated norm $\vert \vert A \vert \vert _2 = (tr A^*A)^{\frac{1}{2}}$

2) The set of Hermitian matrices constitutes a real vector space $\mathbb{H}_n$ in $\mathbb{M}_n$.

3) The subset $\mathbb{P}_n$ consisting of strictly positive matrices is an open subset in $\mathbb{H}_n$. Hence it is a differentiable manifold.

4) The tangent space to $\mathbb{P}_n$ at any of its points A is the space $T_A\mathbb{P}_n = \{A\} \times \mathbb{H}_n$

If this is indeed the proof that I seek then these are the steps that I don't understand:

3) I know that open subsets of manifolds are manifolds, and I am also presuming that "stricly positive" means positive definite. But how do I know that this manifold is differentiable, ie has differentiable transistion maps?

4) This is my main confusion - $\mathbb{H}_n$ is the set of symmetric matrices, but why is the tangent space at any point defined like this? I understand the tangent space to be the collection of velocity vectors through that point. Why can I represent the gradients of all the geodesics that pass through a point like this?

EDIT - UPDATE:

Having been browsing the stack I also came across tangent space clarification where someone is asking why tangent spaces aren't just defined in the same way (ish) as they were in 4) - does anyone know where this definition/way of writing has come from?

$\endgroup$
2
$\begingroup$

3) Any open subset of a smooth manifold is again a smooth manifold (all the charts are simply restrictions of smooth charts to open subsets).

4) The tangent space $T_a V$ to a vector space $V$ (in this case $V=\mathbb{H}(n)$) can be identified with the vector space itself (via the isomorphism which takes an element $v\in V$ to the directional derivative $D_v\vert_a$). Moreover, the tangent space to an open subset of a manifold is isomorphic to the tangent space of the manifold itself. Hence in our case we have $$ T_A\mathbb{P}(n) \cong T_A \mathbb{H}(n) \cong \mathbb{H}(n)\,.$$ For reference I suggest you read the beginning of the chapter "Tangent vectors" in the book "Introduction to Smooth Manifolds" by John M. Lee.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.